There’s no reason to believe that companies running open weights models would use consumer hardware. I can speak on behalf of my employer and say that we’d vastly prefer to have enterprise grade servers running inference on premises.